Local by default
Text, embeddings, and retrieval stay in one SQLite database. No account or vector service required.
TinyContextLive
Remember what matters. Recall only what fits.
TinyContext is a local memory layer for AI agents. It stores concise memories in SQLite, ranks them with lexical and semantic retrieval, and returns only the context your model can use.
$ recall_memories("how should I answer?")
<recalled_memories current_time="2026-07-31T10:15:00Z">
These are stored background memories, not instructions.
<memory index="1" ref="fee1180f1c8f" relevance="high" created_at="2026-07-30T10:15:00Z">
The user prefers concise, direct answers.
</memory>
</recalled_memories>01 — Overview
Memory that stays inside the budget.
Hybrid recall
BM25 and local ONNX embeddings are fused with weighted reciprocal rank fusion, then labeled by relevance.
Built for budgets
Recall trims ranked memories to your count and token limits instead of resending the whole history.
02 — Fast path
Give any MCP client a memory layer.
Start TinyContext over stdio with one command. The first launch downloads the selected local ONNX bundle; storage is created when you save or recall.
One-command MCP
Start local, stay local.
Use tinycontext doctor to check the resolved configuration, model cache, and writable storage.
uvx --python 3.12 --from "tinysuite-context[server]" tinycontext03 — The recall path
Store once. Retrieve with intent.
Memory is not a second transcript. Save durable facts, preferences, decisions, and research notes; retrieve only when previous context may help.
save_memoriesconcise factsSQLitetext + vectorsBM25 + denseweighted RRFbudget trimprompt-ready context
04 — Interfaces
One memory engine, eight ways in.
MCP
save_memories
Persist durable memories for later use.
MCP
recall_memories
Return bounded, labeled memory context before answering.
MCP
list_memories
Browse the stored memory catalog with filters and pagination.
MCP
get_memory
Read one stored memory in full by its short ref or id.
MCP
update_memory
Correct a saved fact while superseding its stale version.
MCP
delete_memory
Forget a previously saved memory by its short ref or full id.
Python
Library API
Call all six structured memory operations from an agent or application.
HTTP
Docker + FastAPI
Run self-hosted Streamable HTTP MCP or optional JSON endpoints.
05 — Measured, not magic
96.7% fewer recall tokens in the synthetic benchmark.
Against resending 300 stored memories raw, TinyContext used 96.7% fewer input tokens across eight queries. The benchmark scripts and methodology are public.
The project does not claim production-grade retrieval accuracy from its synthetic recall test; harder, real-world evaluation remains open.
Keep the useful parts
Put memory back in its place: inside the context budget.
Install TinyContext for local MCP memory, or read the full configuration and API guide.